First, prepare the decision (to ensure good behavior of the
then, a first propagation ensures that, if learning is on,
the unit nogood learnt on failure filters,
the cut is posted before applying the decision to ensure good nogood,
and a second propagation ensures the cut is taken into account
the decision is applied (if learning is on and the decision is refuted,
it is bypassed by the learnt unit nogood),
finally, a fix point is reached.